Research On Image Retrieval Based On Deep Learning And Hashing Technology

With the big data is coming,the number of image in Internet grows explosively every day and a massive information environment has been generated.How to retrieve the images needed accurately and efficiently from a large-scale image dataset has become a very important issue.There are several fundamental problems in current image retrieval methods,such as low expression ability of visual feature,huge computation of high-dimensional vectors' operation and so on.These problems can be solved or relieved by convolutional neural network and hashing technology.Convolutional neural network can learn the intrinsic implications of training images to improve expression ability of visual feature.Hash technology can map high-dimensional deep features into low-dimensional hash codes,which effectively improves retrieval efficiency.The thesis mainly researches on image retrieval based on convolutional neural network and hashing technology,and the contributions are listed as following:(1)Visual features used in state-of-the-art image clustering methods lack of learning,which leads to low expression ability.Furthermore,the efficiency of traditional clustering methods is low for large image dataset.So,a fast image clustering method based on convolutional neural network and binary K-means is proposed.Firstly,a large-scale convolutional neural network is employed to learn the intrinsic implications of training images so as to improve the discrimination and representational power of visual features.Secondly,hashing is applied to map the high-dimensional deep features into low-dimensional hamming space,and multi-index hash table is used to index the initial centers so that the nearest center lookup is carried out efficiently.Finally,image clustering is accomplished efficiently by binary K-means algorithm.Experimental results indicate that the expression ability of visual features is effectively improved and the image clustering performance is substantially boosted compared with state-of-the-art methods.(2)The efficiency of traditional retrieval methods is so low for large image database.In view of this,an image retrieval method based on kernel-based supervised hashing is presented.Firstly,a large convolutional neural network is employed to extract deep feature,and kernel function is used for hash fuction to strengthen the discrimination of non-linear data.Secondly,kernel-based supervised hashing is learned from the high-dimensional visual features,and then maps the latter into compact hamming hash codes.Finally,image retrieval is accomplished in low-dimensional hamming space.Experimental results demonstrate that the retrieval precision and recall are effectively improved,and retrieval time of the new method is much less than that of state-of-the-art methods.(3)There are several images independent with query image in initial ranking list because of noise and semantic gap.In order to further improve the retrieval precision,a structure-based image reranking method is put forward.Firstly,a graph structure is offline built to connect reciprocal neighbor images which can reduce the negative influence on confident samples detection.Secondly,confident samples are detected by finding the weighted maximum density subgraph containing the query image.Finanlly,images in the initial list and confident samples' retrieval list are reranked by evaluating the visual similarity with query.Experimental results indicate that the image retrieval precision is further improved and the new method still performs well in complex environment.